I build the infrastructure that makes AI agents remember, coordinate, execute safely, and improve.
Most AI agent systems are talented amnesiacs. Brilliant in the moment, blank by the next session. They can't remember what they learned, can't build on what came before, and can't coordinate without a human holding the threads.
I build the systems that fix this.
The central nervous system connecting 270+ projects. Seven tiers of memory, seven pluggable executors, OODA composition, and a grammar inspired by the Rig Veda for temporal knowledge annotation.
| Layer | What | How |
|---|---|---|
| Memory | 7-tier system (KV -> Scratchpad -> Vector -> Ephemeral -> Episodic -> Deep/pgvector -> Graph) | Each tier serves a different temporal need |
| Orchestration | OODA loop + AgenticLoop + Dynamic Routing | Agents that observe, orient, decide, act, and reflect |
| Execution | 7 pluggable executors | Subagent, MCP Orchestrator, Content Pipeline, Small LLM, Task Router + 2 API executors |
| Governance | Tiered enforcement + Five-Point Protocol + warrant gate | Not everything deserves to be remembered |
| Knowledge | EBNF grammar (Rig Veda temporal markers) + DIALECTIC engine | Thesis -> antithesis -> synthesis with convergence tracking |
The pipeline that edited my book - 74,000 words, five iterations, score 5.6 -> 9.5 - ran on this.
A default-deny execution boundary for AI agent lanes. It separates orchestration from the act of running commands or local models, applying explicit policy, environment scrubbing, resource limits and structured result contracts.
| Safety surface | What it does |
|---|---|
| Policy validation | Versioned default-deny policy before any process is spawned |
| Environment control | Builds child environments from empty and drops secret-looking variables |
| Process limits | Wall-clock timeout, process-group cleanup, POSIX resource limits and output caps |
| Network posture | Offline by default through environment neutralization and proxy removal |
| Model boundary | Keeps model weights out of git and resolves them from an env-driven runtime cache |
| Local inference | bls infer with llama.cpp-family runners plus a weight-free fake runner for CI |
| Streaming | Additive JSONL transport through bls infer --stream |
| Integration | Stable JSON-in / JSON-out CLI seams for project-broker-loom and other orchestrators |
It is a bounded, policy-driven execution layer, not a kernel or container isolation boundary. The project is public, MIT licensed, protected by a required fail-closed CI gate, and independently developed from project-broker-loom.
AST-driven knowledge graphs over codebases. KuzuDB + hybrid search (BM25 + semantic + graph expansion). 53,000+ entities indexed across 272 projects. 15 MCP tools, 33 REST endpoints, 1,261+ tests.
Explore Intelligence Engine - search, Cypher console, dashboards
Purpose-built Model Context Protocol servers for AI agent workflows.
| Category | Count | Examples |
|---|---|---|
| Orchestration | 4 | SPINE executor adapters, gen-loop scheduler, 8do Ralph Loop |
| Agent Coordination | 4 | agentspool messaging, agent-comm relay, session handover |
| Knowledge & Memory | 5 | Minna (SQLite+FTS5), mem-system (pgvector), context-glue, observation-workbench |
| Research | 3 | research-agent, research-notes, research-log |
| Content & Creative | 4 | content-mcp (41 tools), the-musicologist, content-analyzer, content-extractor |
| Infrastructure | 5 | mcp-server-checker, security-audit, smart-inventory, backup, manual-generator |
| Evaluation | 2 | Evalla (rubric scoring), evaluation-mcp |
| Browser & Visual | 3 | browser-mcp (CDP + grid overlay), tachylite (canvas editor), showcase-mcp |
3 cloud MCP servers live at adaptivearts.ai/mcp - Evalla, Minna Memory, Agent Comm (Bearer auth, rate limiting, audit trail).
Music Video Creator (19 visual styles, beat-sync, genre presets), flow-musical-creation (12-song AI musical), the-musicologist-mcp (style descriptions, Suno/Udio formatting).
A working audio-first pipeline for a daily local morning brief in Västra Götaland, Sweden. The system collects approved public sources, deduplicates and ranks stories, generates a neutral Swedish script, runs editorial QA, creates voice and script-anchored captions, and packages the episode for human review.
It is not an autonomous newsroom and it does not scrape copyrighted articles. The point is a repeatable, editorially controlled production loop: AI-assisted, source-aware, and human-reviewed.
|
|
This is the commercial direction of Adaptivearts.ai in miniature: practical AI systems that connect automation, editorial judgment, safety gates, and reusable production pipelines.
Case study / showcase page coming later.
All self-contained, no API keys needed.
| Showcase | What | Demos |
|---|---|---|
| SPINE Framework | Multi-agent orchestration, 7-tier memory, OODA loops | 11+ interactive demos |
| 8me Learning Platform | Loop orchestration curriculum | 15 progressive labs |
| Intelligence Engine | Code knowledge graphs, hybrid search | Search Picker, Cypher Console |
| Adaptive MCP Orchestrator | Cognitive task dispatcher | Architecture demos |
| Security Audit MCP | 5 security scanners, Docker-isolated | Stack Detector, Secret Scanner |
| Music Video Creator | 19 visual styles, audio analysis | Style Showcase, Beat Effects |
| agentspool | Inter-agent messaging | Network Graph, Message Flow, Radar |
| arbiter | MCP Server Validator - protocol compliance, quality, LLM ergonomics | Profile Simulator, fix packs, checks |
| switchcore | MCP Meta-Router, smart tool routing | Tool cards, architecture |
| vigil | Self-scheduling follow-up server | Tools, check types, notifications |
| spawn | Meta-MCP: builds MCP servers from patterns | Architecture, scoring, pipeline |
By Fredrik Bratten & Sasa Popovic
A 90-day journey from first prompt to production AI systems. 12 chapters across 6 parts. Processed by the SPINE pipeline with AI editorial personas - the same infrastructure described above edited the book that describes it.
9 interactive demos included:
| Demo | What it teaches |
|---|---|
| Prompt Builder | Structured prompt construction |
| Model Selector | Choosing the right model for the task |
| Token Calculator | Understanding token economics |
| Enterprise Flow | Enterprise AI workflow patterns |
| Injection Detection | Prompt injection defense |
| Few-Shot Builder | Few-shot learning patterns |
| MCP Server Setup | Building your first MCP server |
| Five-Point Protocol | Structured execution framework |
| Security Assessment | AI security evaluation |
Independent AI research initiative exploring agentic systems, context engineering, and creative AI pipelines.
Recent applied prototype: an AI-assisted local newsroom pipeline for Västra Götaland, combining source collection, editorial QA, Swedish voice generation, captions, and human review.
- Research & Prototypes - 3 prototype domains with 55+ backing projects
- MCP Server Directory - 5 servers, 49 tools (3 live cloud endpoints)
- Articles & Research - Technical writing on AI architecture, MCP, security
- Context Engineering - Pillar page on context stacks, memory systems, patterns
- AI Agent Architecture - 6-layer architecture reference
20+ years in IT operations, systems architecture, cybersecurity, and DevOps. SOC analysis, XDR implementation, high-availability systems for regulated environments (gaming, finance). Founder of Adaptivearts.ai.
Not Adaptivearts.ai, not an MCP server, not a context-engineering framework. Just my personal music outlet - occasionally the human also emits audio.
Built with SPINE · Adaptivearts.ai